Learning Approximate Preconditions for Methods in Hierarchical Plans Export

@inproceedings{citeulike:688657, author = {Ilghami, Okhtay and Munoz-Avila, Hector and Nau, Dana S. and Aha, David W.}, booktitle = {Proceedings of the 22nd International Conference on Machine Learning}, citeulike-article-id = {688657}, comment = {This paper describes an extension to earlier work that learned the preconditions of methods in an HTN by observing plan traces and using the Candidate Elimination algorithm to find the single set of preconditions that is always satisfied when a method is used. These plan traces include both the methods actually chosen and any other methods that could have been chosen, which is strange, and are generated by the SHOP planner with full knowledge of the domain. The problem with the earlier work is that it required very many plan traces to converge (completely determine the preconditions of all methods). This work concludes that the convergence problem is due to the fact that some situations are very rare in the training set, and that thus the information learned may be useful before learning is complete. Determining whether or not to use a method uses nodes in a version space, which is part of the candidate elimination algorithm, to vote on whether or not the current state fulfills the precondtions associated with that node. If the acceptance ratio is greater than some threshold, the method is considered usable. To prevent unnecessary branching during the plan search process, a beam size is used such that the number of methods considered for a task cannot exceed some set limit. This algorithm is used to solve problems in two domains, Blocks World and Noncombatant Evacuation Operations. The algorithm works quite well in both cases, but different parameterizations are most effective for the two domains. The measure of quality is precision: the ratio of correct solutions to problems presented. It is not clear what the algorithm does when two or more plans are possibly correct but it lacks the training data to determine whether or not they are actually correct. Presumably, one is chosen and if it is not correct the planner is counted as failing. I will need to research the Candidate Elimination algorithm in order to fully understand this work, but it seems like an incremental improvement over the existing system. Precision is significantly better, but the idea of using incomplete information seems rather obvious. Furthermore, the idea that methods could be known but not their preconditions seems unlikely to actually occur.}, keywords = {htns, learning}, posted-at = {2006-06-07 17:40:20}, priority = {0}, title = {Learning Approximate Preconditions for Methods in Hierarchical Plans}, year = {2005} }

TY - CONF ID - citeulike:688657 L3 - citeulike-article-id:688657 TI - Learning Approximate Preconditions for Methods in Hierarchical Plans T2 - Proceedings of the 22nd International Conference on Machine Learning KW - htns KW - learning N1 - This paper describes an extension to earlier work that learned the preconditions of methods in an HTN by observing plan traces and using the Candidate Elimination algorithm to find the single set of preconditions that is always satisfied when a method is used. These plan traces include both the methods actually chosen and any other methods that could have been chosen, which is strange, and are generated by the SHOP planner with full knowledge of the domain. The problem with the earlier work is that it required very many plan traces to converge (completely determine the preconditions of all methods). This work concludes that the convergence problem is due to the fact that some situations are very rare in the training set, and that thus the information learned may be useful before learning is complete. Determining whether or not to use a method uses nodes in a version space, which is part of the candidate elimination algorithm, to vote on whether or not the current state fulfills the precondtions associated with that node. If the acceptance ratio is greater than some threshold, the method is considered usable. To prevent unnecessary branching during the plan search process, a beam size is used such that the number of methods considered for a task cannot exceed some set limit. This algorithm is used to solve problems in two domains, Blocks World and Noncombatant Evacuation Operations. The algorithm works quite well in both cases, but different parameterizations are most effective for the two domains. The measure of quality is precision: the ratio of correct solutions to problems presented. It is not clear what the algorithm does when two or more plans are possibly correct but it lacks the training data to determine whether or not they are actually correct. Presumably, one is chosen and if it is not correct the planner is counted as failing. I will need to research the Candidate Elimination algorithm in order to fully understand this work, but it seems like an incremental improvement over the existing system. Precision is significantly better, but the idea of using incomplete information seems rather obvious. Furthermore, the idea that methods could be known but not their preconditions seems unlikely to actually occur. AU - Ilghami, Okhtay AU - Munoz-Avila, Hector AU - Nau, Dana AU - Aha, David PY - 2005/// ER -